Predicting disease states using AI image recognition
The Francis Crick Institute
We developed technology that can classify four subtypes of Parkinson’s disease with up to 95% accuracy. Identifying disease subtypes is the first step towards developing personalised treatments based on an individual’s condition.
Level of accuracy reached when classifying Parkinson’s
Background
Currently, there are no approaches to tell apart the subtypes of Parkinson’s disease via their distinct molecular mechanisms. As a consequence, diagnosis and treatments made available to patients are not personalised. This motivated The Francis Crick Institute to develop a computer vision model to diagnose Parkinson’s disease and, crucially, its subtypes.
“This technique can be carried out automatically on a variety of images, which would be hard and time-consuming for a human to assess and also prone to error.”
Giulia Vecchi, Senior Data Scientist
Faculty
Solution
We created a computer vision model, an AI application that analysed microscope images of neurons to distinguish between healthy tissues and different subtypes of Parkinson’s disease. The AI can predict the subtypes in new images with up to 95% accuracy, even highlighting the areas in the images that drove its decision.
Impact
The programme rapidly accelerated data science capabilities at Francis Crick Institute, saving time and resources while helping to discover new ways to understand, diagnose and treat Parkinson’s. This research was published in a high-impact journal, Nature Machine Learning, highlighting its significance in advancing new capabilities in the field.
“The capability to accurately predict a disease mechanism in a brain cell unlocks the potential to design different treatments for different disease subtypes - an exciting approach for treating highly variable diseases in future.”